Image super-resolution with sparse neighbor embedding.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Neighbor-embedding-based (NE) super-resolution (SR) algorithms traditionally use two separate processes: neighbor search and weight determination.
• These independent processes are suboptimal, potentially limiting reconstruction quality and efficiency.

Purpose of the Study:

• To propose a unified sparse neighbor selection scheme for SR reconstruction.
• To enhance the efficiency and quality of high-resolution (HR) image synthesis from low-resolution (LR) inputs.

Main Methods:

• A sparse neighbor selection scheme is developed, using an extended Robust-SL0 algorithm for simultaneous neighbor identification and weight solving.
• Histograms of Oriented Gradients (HoG) are employed for clustering LR patches based on local geometric structures.
• k-nearest neighbors (k-NN) are adaptively selected from clustered subsets using HoG information.

Main Results:

• The proposed method integrates neighbor selection and weight determination into a single, optimized process.
• Utilizing HoG for clustering and adaptive k-NN selection significantly improves the speed of HR image synthesis.
• Experimental results demonstrate competitive SR quality compared to state-of-the-art methods.

Conclusions:

• The proposed sparse neighbor selection scheme offers a more optimal approach to SR reconstruction.
• Integrating local structural information via HoG enhances both the speed and quality of SR image synthesis.
• This method presents a promising advancement in the field of image super-resolution.